Harbour & Pine Distribution
1. Summary
Workflow picked: matching supplier invoices to purchase orders and delivery notes.
What the prototype does: reads each supplier invoice from the accounts inbox, finds the matching purchase order and delivery note, and flags any difference in price, quantity or item. A person in the finance team approves every match before anything is posted.
Recommendation: move to a Custom Workflow Build. The cost case below pays back in 5 to 10 months on cautious assumptions.
2. Where the time goes (week one)
We mapped five recurring workflows with the people who do them. Hours are the team's own monthly estimates.
| Workflow | People | Hours a month | Decision |
|---|---|---|---|
| Supplier invoice matching | 3 | 70 | Picked |
| Customer order emails keyed into the ERP | 4 | 40 | Later: customer-facing, needs more controls first |
| Delivery exception follow-up | 2 | 30 | Later: depends on carrier data we couldn't access |
| Weekly sales report | 1 | 24 | Ruled out: the existing reporting tool can do this without AI |
| New supplier set-up | 1 | 10 | Ruled out: too little time to justify a build |
Why invoice matching first: it takes the most hours, the rules are already written down in the finance manual, the data is already digital, and mistakes cost money through late-payment fees and overpayments.
3. The cost case
| Item | Figure | Assumption |
|---|---|---|
| Time spent today | 70 hours a month | Team estimate, checked against one week of timesheets |
| Cost of that time | $2,100 a month | $30 an hour, fully loaded, from the finance team |
| Time saved after a Build | 35 to 50 hours a month | Based on the prototype test below; people still review every flagged invoice |
| Monthly saving | $1,050 to $1,500 | Hours saved multiplied by $30 |
| Estimated Build cost | $8,000 to $10,000 | Includes the ERP connection and four weeks of build and testing |
| Payback | 5 to 10 months | Build cost divided by monthly saving |
4. The prototype (week two)
We built the prototype in Harbour & Pine's own n8n account and tested it on 200 past invoices that the finance team had already matched by hand.
- 162 invoices matched automatically, and the finance team agreed with every one.
- 38 invoices were flagged for a person: 29 had a real price or quantity difference, and 9 were scanned too poorly to read.
- It posted nothing to the ERP and paid nothing. It only prepared matches for approval.
5. Rules the prototype follows
- It never approves, posts or pays an invoice. A named person in finance does.
- Bank details on invoices are masked before any text reaches the AI model.
- Any invoice it can't read clearly goes to a person, not to a guess.
- Every run is logged: what came in, what it matched, and who approved it.
6. Plan to run it
- Build scope: connect to the ERP, add the approval screen, handle credit notes, and train the finance team. About four weeks.
- Owner: the accounts payable lead, with the Head of Finance as sponsor.
- Success measures after 90 days: hours spent on matching, share of invoices matched without changes, and time from invoice received to approved.
- Go or no-go check: at day 30, compare hours and error rates with the baseline above before rolling out to all suppliers.
7. Risks we'd watch
- Poor scans from a few suppliers. Fix: ask those suppliers for PDF invoices.
- Purchase orders raised after the invoice arrives. Fix: hold those invoices for a person rather than forcing a match.
- The saving depends on the team trusting the matches. Fix: run side by side for the first month.
Want a report like this on your own workflow? Founding clients pay $3,200 (SAR 12,000) instead of $4,500 (SAR 16,900) for a Workflow Discovery Sprint, until 27 April 2027. 3 places.
Book a Sprint call